SISOFT SANDRA 2010 was just released, and the big news is that it now has OpenCL based GPGPU benchmarks. They look to be a fairly well rounded and test both compute capabilities and I/O.
The new Sandra 2010 benchmark suite has four modules that will be relevant to those curious about GPGPU performance, processing, bandwidth, video rendering and video memory. It also supports ATI’s Stream, Nvidia’s CUDA, OpenCL and DX11 compute shaders so you can test just about any major card feature through any major API you want. There is a good writeup of the updates here.
There are four major points of a GPU to test, PCIe bandwidth utilization, video memory bandwidth, single-precision floating-point (SP FP) and double-precision floating-point (DP FP) number crunching. Each of the above tests singles out a specific feature that needs testing, so if you are overly concerned with a single property, you can look at the subtest and get a more directly relevant answer than the general system score. If you want more on the OpenCL support, read this.
People who have tested early versions of Sandra 2010 tell SemiAccurate that the numbers are pretty close to parity between the major native APIs, Stream, CUDA and OpenCL, something that is quite impressive given the relative immaturity of OpenCL.
SiSoft published some basic benchmarks comparing both CPUs and GPUs. Its bandwidth numbers are here, and its FP numbers are here. Needless to say, GPUs crush most modern CPUs, but that was expected. Once you throw a single serial step in, the GPUs don’t look so good, but this is a relative test of their strengths.
The summary of the results is that in raw performance, the ATI 5870 blows the others out of the water in quite abusive ways. The results get even more abusive when you consider that the ATI 5870 running on the beta OpenCL from AMD beats an Nvidia GTX295 running CUDA by 2.7 times. Worst case ATI 5870 beats the best case dual card from Nvidia by more than twice.
In the end, we finally have what looks to be a decent, independent and repeatable GPGPU benchmark. It tests not only the aggregate performance of the parts, but also the four most relevant statistics. If you are interested, you can download it from SiSoft, and best of all, the Lite version, with some restrictions, is free.S|A
Latest posts by Charlie Demerjian (see all)
- Globalfoundries 7nm process isn’t even close to the name - Sep 26, 2016
- ARM upgrades realtime offerings to v8-R and adds Cortex-R52 - Sep 21, 2016
- Everspin and Globalfoundries team up for embedded ST-MRAM - Sep 15, 2016
- Intel’s Xpoint is pretty much broken - Sep 12, 2016
- ARM adds 2048-bit vectors to v8A with SVE - Sep 7, 2016